Sovereign deployment could change the cost of supplying clinical AI, but it does not establish that customers will pay less. A supplier might reduce spending on external model services while taking on more infrastructure, engineering and maintenance work. It might pass a saving to customers, invest it elsewhere or retain it as margin.
For a clinic, the more useful question is what it costs to obtain an acceptable completed workflow. A low processing price can coexist with substantial review and correction work.
The public material examined for Tandem Sovereign Health does not provide a deployment-specific quotation on which to base a price comparison. The numerical example below is therefore entirely hypothetical. It is not a Tandem estimate, competitor quote or observed performance result.
Separate model cost from the service being purchased
The underlying model is one component of a clinical product. The service may also require capture, storage, evidence access, integrations, clinical evaluation, security, customer support and incident investigation.
Operating model weights directly does not make those activities disappear. A hosted external model and supplier-operated infrastructure allocate the work differently.
A buyer should consequently distinguish a price for computation from a price for a supported clinical application. The phrase "open-weight" identifies an access and deployment property, not a guarantee of free operation or unrestricted licensing.
The AI harness explainer describes why the surrounding system matters. Its cost matters for the same reason: the customer relies on the assembled product rather than a model file in isolation.
Compare operating models at the right workload
A usage-based model service can suit variable demand because charges track activity. Dedicated capacity can create a different pattern, with more predictable resources but a requirement to pay for capacity that may not always be used.
Those are economic mechanisms to model, not universal claims that one arrangement is cheaper. Traffic peaks, latency requirements, support obligations and the ability to share infrastructure affect the comparison.
The government's cloud-spending guidance treats financial management and operational choices together. For healthcare, the analysis should extend to the human work required to use the resulting output responsibly.
An organisation should not accept a calculation that assumes maximum utilisation all day when its actual demand is concentrated into clinics and leaves long periods with little activity.
Define the unit before dividing the costs
A generated note is not necessarily a usable note. A reviewed note is not necessarily a successfully filed record. A completed consultation may also require correspondence and follow-up outside the documentation task.
For the example below, the unit is deliberately narrow: one completed documentation episode, including review and any additional manual rewriting required. It is not the cost of providing the whole consultation.
Both options support 1,000 episodes per month. In each, 950 generated outputs are assumed usable after ordinary review and 50 require an additional five-minute manual rewrite. This is an invented operational assumption, not a measured 95% clinical-accuracy rate.
Every output is reviewed. The manual rewrite is additional work for the affected fifty episodes, so the calculation does not silently exempt unsuccessful outputs from review.
A worked example: the cheaper service invoice loses overall
Assume clinician time is valued at GBP 60 an hour, equivalent to GBP 1 a minute. That valuation represents an economic use of time, not an automatic reduction in salary expenditure.
| Monthly documentation cost | Option A | Option B |
|---|---|---|
| Service charge | GBP 900.00 | GBP 600.00 |
| Local assurance and support allocation | GBP 250.00 | GBP 450.00 |
| Initial implementation spread over twelve months | GBP 100.00 | GBP 300.00 |
| Review of all 1,000 outputs | GBP 333.33 | GBP 583.33 |
| Additional manual rewriting of 50 outputs | GBP 250.00 | GBP 250.00 |
| Total modelled monthly cost | GBP 1,833.33 | GBP 2,183.33 |
| Cost per completed documentation episode | GBP 1.83 | GBP 2.18 |
Option A assumes twenty seconds of review per output. Option B assumes thirty-five seconds. The implementation rows allocate hypothetical upfront costs of GBP 1,200 and GBP 3,600 respectively over twelve months; they do not describe monthly implementation invoices.
The arithmetic for A's review is 1,000 x 20 seconds / 60, valued at GBP 1 per minute: GBP 333.33. For B it is 1,000 x 35 / 60: GBP 583.33. Manual rewriting costs 50 x 5 x GBP 1, or GBP 250 for each.
Option B's service invoice is GBP 300 lower, yet its total modelled cost is GBP 350 higher. The conclusion follows from the stated assumptions, not from any finding that sovereign or open-weight products inherently need longer review.
Change the assumptions before trusting the conclusion
The example is useful only if the reader can challenge it.
An extra fifteen seconds of review across 1,000 episodes represents 250 minutes, valued here at GBP 250. That difference is large enough to alter an apparently attractive licence saving.
Now change the volume. An implementation cost spread across more episodes has a smaller per-episode effect. A fixed infrastructure commitment may become more attractive with sustained demand, while a lightly used deployment may struggle to justify it.
Change the rewrite rate as well. A product can be inexpensive to review most of the time but generate occasional exceptions requiring substantial work. Those exceptions should not disappear into a favourable average.
A real business case should use local measurements, uncertainty ranges and clinically appropriate quality criteria. It should not optimise a cost model by assuming that staff can safely omit necessary review.
Be explicit about what the calculation leaves out
The example excludes the ordinary clinical consultation, common record-system costs, taxes and financing effects. It does not value patient outcomes or claim that a correct documentation episode settles every care task.
Additional costs may be needed for a particular deployment. Conversely, some local support work may already be funded and should not be counted twice.
The purpose is to make the boundary inspectable. A number labelled "cost per consultation" would be misleading if it actually included only the AI supplier's inference bill.
The procurement questions provide a route to collecting the information needed for a more complete organisational comparison.
Released time and cash savings are different benefits
Reducing documentation work can create time for other tasks, reduce overtime or improve the pace of a clinic. Whether it saves cash depends on what the organisation actually changes.
A salaried clinician spending fewer minutes correcting notes does not automatically reduce the monthly payroll. The gain may instead be capacity, working conditions or the ability to complete necessary work within contracted time.
Those benefits can still justify investment. They simply require their own description and evidence.
Similarly, increasing appointment capacity may reveal more demand and generate additional investigation or follow-up. It should not be described as a cost reduction solely because each note is produced faster.
Where a supplier's saving might go
A supplier could lower prices, increase margins, improve service availability, invest in evaluations or add capabilities. These possibilities are not mutually exclusive.
A larger margin is not evidence of wrongdoing. Nor does a lower price demonstrate that the service has cut essential assurance. The relevant question is what the actual contract and product deliver.
Ask for the pricing basis, volume assumptions, included support, implementation charges, change costs and exit terms. Establish whether a more dedicated deployment introduces minimum commitments or resources that an individual subscription does not require.
Sovereignty can be a worthwhile purchasing objective even when it costs more. The honest business case identifies the control gained and the price of obtaining it, rather than inventing an infrastructure saving to justify a decision already made.
Frequently asked questions
Are open-weight clinical AI models free to operate?
No. Access to weights does not remove infrastructure, engineering, licensing, evaluation or support costs. The finished clinical product has additional responsibilities beyond running the model.
Are the prices in the worked example Tandem quotations?
No. Every value and performance assumption is hypothetical. The example demonstrates a calculation method and should be replaced with the buyer's actual quotations and measurements.
Does saving clinician time automatically save the same amount of money?
No. Valuing time helps compare resource use, but cash savings require a corresponding change in expenditure. Released capacity or reduced workload can be worthwhile benefits without being payroll savings.
Examine the full workflow behind a clinical AI business case